1 Acoustical characteristics based predictive diagnostics of individual dry pump for semiconductor manufacturing process Kyuho Lee 1 , Soogab Lee 1 , Jong-Yeon Lim 2 and Wan-Sup Cheung 2 , 1 Department of Mechanical and Aerospace Engineering, Seoul National University, 599 Gwanak-ro, Gwanak, Seoul, Republic of Korea 2 Acoustics & Vibration Group, KRISS 1 Doryong-Dong, YUSONG, DAEJON 305-340, Republic of KOREA ABSTRACT This paper addressed the acoustical characteristics of dry vacuum pumps and predictive diagnostics of dry vacuum pumps with these characteristics. The dry pump system designed for the semiconductor manufacturing is used to maintain the cleanliness by exhaust the purge gas. The semiconductor process can be divided two separable state segments, the one is gas loaded state batch when pump exhaust the purge gas and the other is idle state batch when the valve is closed during pump operation. As the physical attributions of gas loaded state and idle state are different, the corresponding noise characteristic is different, too. With the Linearized Adaptive Parameter Model (APM), the parameters indicate each batch can be obtained. These parameters contain the characteristics of pump noise and can be used for predictive diagnostic of individual dry pump by observing the trend of parameters. 1. INTRODUCTION In semiconductor manufacturing process, the need for predictive diagnostic technique for dry vacuum pump has been one of the “hot” technical issues since R. Bahren and M. Kuhn [1] pointed out its significance. One of main applications of the dry vacuum pump system has focused on the semiconductor manufacturing processes that require much improved cleanliness [2]. So, diagnostic technique for dry vacuum pump system has strong relationship with an error rate of semiconductor. The test result, carried out in the Centre of Vacuum Technologies of KRISS, indicate that state variables such as exhaust pressure and supply currents to booster and dry pump motor are monitored only as static properties. Lim et al.[3] proposed the use of vibration accelerometers to monitor the dynamic running conditions of the gears and bearing of vacuum pumps, including the unbalance of rotors. A common idea for predictive diagnostics, as described by Robert et al.[4], is to extract extraordinary features from the recorded signals of state variables using multiple principal component analysis (PCA). These principal components are converted to one value, Hotelling’s T 2 . However, much difficulty in using PCA is encountered when the sizes of the collected batch data are different each others. Unfortunately, the semiconductor process period is time-varying, so the sizes of the collected state variables batch are different. D. Sung [5] reported that the dynamic time warping algorithm[6,7] worked well for predictive diagnosis of dry vacuum pumps by warping collected state variables batch data. To overcome that DTW take number of computation resources, W. Chueng [8] proposed linear adaptive parameter modeling (APM) algorithm and K. Lee [9] applied APM to pump diagnosis. To use adaptive parameter modeling, the characteristics of collected state variable batch data have to be considered first. K. Lee [9] divided semiconductor manufacturing processes into 1 bukha16@snu.ac.kr (the first author)